# Second Brain

*Last synthesized: 2026-10-10 | 44 files | 6 concept pages | offline, zero tokens*

> Raw sources -> readmenator wiki -> links (Karpathy LLM Wiki Pattern, deterministic).
> Start here, then open one community page. Prefer grep over full reads.

## Vault Overview

The codebase centres on `model.py`, `__init__.py`, `train.py`. Architecturally it is 6 layers, dominant utility (36 files) across 6 import-based communities. Recorded risk surface: 0 security findings and 0 dependency cycles.

Surprising tissue lives between topogpt3: model, eval: harness, eval: governor: 8 extracted cross-community imports and 12 inferred bridges. Follow `connections.json` sorted by strength before refactoring.

Open work clusters around documentation (82% file coverage), 0 security findings, 20 taint paths, and 5 suggested exploration questions in `queries.md`.

## Stats

| Metric | Value |
|--------|-------|
| Files | 44 |
| Symbols | 824 |
| Resolved imports | 66 |
| Languages | py, sh |
| Communities | 6 |
| Doc coverage | 82% (36/44 files) |
| Security findings | 0 |
| Estimated read cost | ~26833 tokens (chars/4, offline so $0) |

## Reading Order

1. Skim Stats and God Nodes below for blast radius.
2. Open the largest community page first, then follow Connections.
3. Use `queries.md` for the next question; log the answer there.

```
grep -rn '<keyword>' index.md community_*.md
readmenator query "<question>" --target TopoExploit
```

## Concept Wiki

- [topogpt3: model (12 files, cohesion 0.64)](./community_0_topogpt3_model.md)
- [eval: harness (7 files, cohesion 0.64)](./community_1_eval_harness.md)
- [eval: governor (6 files, cohesion 0.31)](./community_2_eval_governor.md)
- [topogpt3: inference_hrm (5 files, cohesion 0.42)](./community_3_topogpt3_inference_hrm.md)
- [topogpt3: jlens (4 files, cohesion 0.45)](./community_4_topogpt3_jlens.md)
- [orphans (10 files, cohesion 0.00)](./community_5_orphans.md)

## God Nodes

| File | Score |
|------|-------|
| `topogpt3/model.py` | 48.7 |
| `topogpt3/__init__.py` | 30.0 |
| `topogpt3/train.py` | 24.4 |
| `topogpt3/lens_model.py` | 14.9 |
| `topogpt3/inference_hrm.py` | 13.6 |

## Strongest Connections

- 0 -> 2: depends_on (strength 0.9, EXTRACTED)
- 1 -> 2: depends_on (strength 0.9, EXTRACTED)
- 1 -> 0: depends_on (strength 0.9, EXTRACTED)
- 4 -> 0: depends_on (strength 0.9, EXTRACTED)
- 2 -> 3: depends_on (strength 0.9, EXTRACTED)
- 2 -> 4: depends_on (strength 0.9, EXTRACTED)
- 3 -> 4: depends_on (strength 0.9, EXTRACTED)
- 3 -> 0: depends_on (strength 0.9, EXTRACTED)
- 2 -> 1: bridges (strength 0.5, INFERRED)
- 2 -> 0: bridges (strength 0.5, INFERRED)

## Navigation Tips

- Obsidian Graph View works: every community page links back here.
- `connections.json` is machine-readable for GraphRAG pipelines.
- `REPORT.md` states what was extracted vs inferred and current limits.
- Regenerate offline: `readmenator . --rebuild` (no network, no tokens).